Synopsis

Keywords: Quantitative Imaging, DSC & DCE Perfusion

In DCE-MRI, advanced pharmacokinetic models, such as the 2CX, TH and ATH models, provide a more complete set of pharmacokinetic parameters than the commonly used simpler models, such as the Tofts and Patlak models. However, their use requires higher temporal resolution and signal-to-noise ratio. Hence, advanced-model DCE-MRI allows for imaging of only a few slices. We present an approach making 3D advanced-model DCE-MRI possible and evaluate it under the low-signal-to-noise-ratio conditions of small-animal MRI. Our methodology is based on compressed sensing and spatially regularized fitting of the pharmacokinetic model. The approach is evaluated on simulated and real datasets.

INTRODUCTION

Use of advanced pharmacokinetic (PK) models¹ (e.g. 2CX, TH, ATH) in Dynamic Contrast-Enhanced (DCE) MRI can provide a more complete set of PK parameters than the common models¹ (e.g. Tofts, Patlak), especially by estimating blood flow and permeability surface area instead of their compound parameter Ktrans. However, advanced PK models require high temporal resolution of the acquisition (approx. 1s) and high signal-to-noise ratio (SNR) (especially challenging in small-animal DCE-MRI). This allows for acquisition of one or only a few slices. We show the feasibility of 3D advanced-model DCE-MRI of small animals by means of compressed sensing and spatially regularized perfusion analysis.

METHODS

We used our implementation of the radial golden-angle stack-of-stars (GASOS) T₁-weighted spoiled gradient-echo acquisition² to achieve the incoherent-sampling condition of compressed sensing. From the acquired data, we reconstructed image sequences by the L+S algorithm with limited differences³, minimizing the time-domain starcase artifacts. Subsequently, we converted the image-sequences to contrast-agent concentration curves using T₁₀ mapping from multi-flip-angle precontrast recordings. Finally, we estimated PK maps with spatially regularized PK-model fitting⁴. In contrast to standard voxel-by-voxel approach, the spatially regularized version fits the PK model simultaneously to all voxels of a slice while inducing edge-preserving smoothness (formulated as total-variation regularization) of the resulting PK maps. The tissue homogenity (TH) PK model was used, with the fitting procedure formulated in the Fourier domain (including estimation of the bolus arrival time), assuming a population-based arterial input function⁵.

To evaluate the proposed approach with respect to the ground truth and under ideal conditions, we generated simulated k-space data. The GASOS acquisition was simulated for the same acquisition parameters as used for real data described below. We added realistic Gaussian noise of approx. the same level as in real data. The concentration curves were generated using the TH model. Based on a real DCE-MRI recording of a rat with a glioma, we manually segmented out 53 regions and assigned each region a median of its estimated voxel-based PK parameters as the ground-truth value. Constant T₁₀ and T₂₀ were assumed. Multi-flip-angle precontrast sequences were simulated as well.

Furthermore, we evaluated our method on recordings from 6 tumor bearing SHO mice (MiaPaCa-2 luc SPX2 ovarian tumor cells; approved by the National Animal Research Authority), imaged with single-slice 2D GA DCE and with the 3D GASOS DCE (2 boli of Magnevist, Bayer HealthCare, Germany, separated by 45 min). We used the 2D recording (image sequence reconstructed using NUFFT, no spatial regularization in PK model fitting) as a reference. For comparison of our 3D DCE results with this 2D reference, we selected manually the corresponding slice of the 3D data. The details of our acquisition procedure were as follows: 9.4T scanner Bruker Biospin (Ettlingen, Germany), 4-channel surface mouse brain coil, 2D DCE: TR/TE = 15/1.5ms, FA = 20°, 1 axial slice, 128 samples/projection, acquisition time 15 min, multi-TR precontrast scans with TR=15, 30, 50, 100, 250, 500ms; 3D DCE: TR/TE = 5.5/1.4ms, FA = 15°, 25 slices, 128 samples/projection, acquisition time 15 min, multi-flip-angle precontrast scans with FA=5, 10, 15, 25, 30°.

RESULTS

As shown in Figs.1,3, the PK maps of the 3D data were clearly unacceptable when using only the NUFFT reconstruction. A substantial improvement was achieved by the L+S reconstruction and a slight further improvement was gained by the subsequent regularized PK-model fitting. This agreed well with the quantitative comparison of the mentioned 3D approaches with the reference 2D results (Fig.4) using the mean voxel-wise within-subject coefficient of variation (wCV)⁶ of the PK-parameter estimates. Fig.2 shows the quality of images and time curves reconstructed from the 3D data with L+S, comparable with the reference 2D data.

DISCUSSION

The most important step towards feasible 3D DCE-MRI was the GASOS acquisition combined with the L+S reconstruction. The subsequent spatial regularization of PK-model fitting was helpful mainly for estimates of Fp, Tc and E, which are the most difficult PK parameters to estimate. In evaluation on the real data, the comparison of the PK maps obtained from the reference 2D data and the corresponding slice of the 3D datasets was problematic because of the impossibility of choosing exactly the same slice, due to animal motion during the examination. This was the reason for some mismatch between the 2D and 3D PK-parameter maps.

CONCLUSION

Advanced-model 3D DCE-MRI in small animals is feasible for tumor applications. Compared to the available simple-model DCE-MRI, or single-slice advanced-model DCE-MRI, the proposed approach can provide a broad set of PK parameters within the whole tumor volume, allowing for studies of tumor heterogeneity.

Acknowledgements

Supported by Czech Science Foundation (GA22-10953S) and Ministry of Education, Youth and Sports of the Czech Republic (LM2018129).